This action might not be possible to undo. Are you sure you want to continue?
Fluid Dynamics (1998) 10: 213–237
Theoretical and Computational
© SpringerVerlag 1998
Fluid Dynamics
Optimum Aerodynamic Design Using the
Navier–Stokes Equations
1
A. Jameson and L. Martinelli
Department of Mechanical and Aerospace Engineering, Princeton University,
Princeton, NJ 08544, U.S.A.
N.A. Pierce
Oxford University Computing Laboratory, Numerical Analysis Group,
Oxford OX1 3QD England
Communicated by M.Y. Hussaini
Received 5 February 1997 and accepted 30 May 1997
Abstract. This paper describes the formulation of optimization techniques based on control theory for
aerodynamic shape design in viscous compressible ﬂow, modeled by the Navier–Stokes equations. It extends
previous work on optimization for inviscid ﬂow. The theory is applied to a system deﬁned by the partial
differential equations of the ﬂow, with the boundary shape acting as the control. The Fr´ echet derivative of the
cost function is determined via the solution of an adjoint partial differential equation, and the boundary shape
is then modiﬁed in a direction of descent. This process is repeated until an optimum solution is approached.
Each design cycle requires the numerical solution of both the ﬂow and the adjoint equations, leading to a
computational cost roughly equal to the cost of two ﬂow solutions. The cost is kept low by using multigrid
techniques, in conjunction with preconditioning to accelerate the convergence of the solutions. The power of
the method is illustrated by designs of wings and wing–body combinations for long range transport aircraft.
Satisfactory designs are usually obtained with 20–40 design cycles.
1. Introduction
This paper, which is dedicated to Sir James Lighthill, is focused on the problem of aerodynamic design.
Here, as in so many other branches of ﬂuid mechanics and applied mathematics, Lighthill has made a
seminal contribution through his early demonstration of a solution of the inverse problem for airfoil design
in potential ﬂow [1].
The evolution of computational ﬂuid dynamics during the last three decades has made possible the
rapid evaluation of alternative designs by computational simulation, eliminating the need to build numerous
models for wind tunnel testing, which is used primarily to conﬁrm the performance of the ﬁnal design, and
to provide a complete database for the full ﬂight envelope. The designer still needs some guiding principle
to distinguish a good design out of an inﬁnite number of possible variations, since it is not at all likely
that a truly optimum design can be found by a trial and error process. This motivates the use of numerical
optimization procedures in conjunction with computational ﬂow simulations.
1
This work has beneﬁted fromthe generous support of AFOSRunder Grant No. AFOSR910391, DOD/URI/ONR/ARPAN00014
92J1796, the NASA–IBM Cooperative Research Agreement, EPSRC and the Rhodes Trust.
213
214 A. Jameson, L. Martinelli, and N.A. Pierce
Early investigations into aerodynamic optimization relied on direct evaluation of the inﬂuence of each
design variable. This dependence was estimated by separately varying each design parameter and recalcu
lating the ﬂow. The computational cost of this method is proportional to the number of design variables and
consequently becomes prohibitive as the number of design parameters is increased.
An alternative approach to design relies on the fact that experienced designers generally have an intuitive
feel for the kind of pressure distribution that will provide the desired aerodynamic performance. This
motivates the introduction of inverse problems in which the shape corresponding to a speciﬁed pressure
distribution is to be determined. A complete analysis of the inverse problem for airfoils in two dimensional
potential ﬂow was given by Lighthill [1], who obtained a solution by conformally mapping the proﬁle to a
unit circle. The speed over the proﬁle is
q =
1
h
∇ϕ,
where ϕ is the potential, which is known for the circle, while h is the modulus of the mapping function. The
surface value of h can be obtained by setting q = q
d
, where q
d
is the desired speed, and since the mapping
function is analytic, it is uniquely determined by the value of hon the boundary. Lighthill’s analysis highlights
the fact that a physically realizable shape may not exist unless the prescribed pressure distribution satisﬁes
certain constraints. In fact a solution exists for a given speed q
∞
at inﬁnity only if
1
2π
_
q
d
dθ = q
∞
,
where θ is the polar angle around the circle, and there are additional constraints on q
d
if the proﬁle is to be
closed.
In the more general case of threedimensional viscous compressible ﬂow, the constraints which must be
satisﬁed by a realizable target pressure distribution are not known, and attempts to enforce an unrealizable
pressure distribution as a boundary condition can lead to an illposed problem. The problems of optimal and
inverse design can both be systematically treated within the mathematical theory for the control of systems
governed by partial differential equations [2] by regarding the design problem as a control problem in which
the control is the shape of the boundary. The inverse problem then becomes a special case of the optimal
design problem in which the shape changes are driven by the discrepancy between the current and target
pressure distributions.
The control theory approach to optimal aerodynamic design, in which shape changes are based on gradient
information obtained by solution of an adjoint problem, was ﬁrst applied to transonic ﬂow by Jameson [3],
[4]. He formulated the method for inviscid compressible ﬂows with shocks governed by both the potential
equation and the Euler equations [3], [5], [6]. With this approach, the cost of a design cycle is independent
of the number of design variables. When applied to the design of the airfoils in compressible potential
ﬂow using conformal mapping to transform the computational domain to a unit disk, it leads to a natural
generalization of Lighthill’s method. The effects of compressibility are represented by an additional term in
the modiﬁcation of the mapping function which tends to zero as the Mach number tends to zero [3], [5]. More
recently, the method has been employed for wing design in the context of complex aircraft conﬁgurations
[7], [8], using a grid perturbation technique to accommodate the geometry modiﬁcations.
Pironneau had earlier initiated studies of the use of control theory for optimum shape design of systems
governed by elliptic equations [9], [10]. Ta’asan et al. have proposed a one shot approach in which the
constraint represented by the ﬂow equations need only be satisﬁed by the ﬁnal converged design solution
[11]. Adjoint methods have also been used by Baysal and Eleshaky [12], by Cabuk and Modi [13], [14], and
by Desai and Ito [15].
The objective of the present work is the extension of adjoint methods for optimal aerodynamic design
to ﬂows governed by the compressible Navier–Stokes equations. While inviscid formulations have proven
useful for the design of transonic wings at cruise conditions, the inclusion of boundary layer displacement
effects with viscous design provides increased realismand alleviates shocks that would otherwise formin the
viscous solution over the ﬁnal inviscid design. Accurate resolution of viscous effects such as separation and
shock/boundary layer interaction is also essential for optimal design encompassing offdesign conditions
and highlift conﬁgurations.
The computational costs of viscous design are at least an order of magnitude greater than for design using
the Euler equations for several reasons. First, the number of mesh points must be increased by a factor of
Optimum Aerodynamic Design Using the Navier–Stokes Equations 215
two or more to resolve the boundary layer. Second, there is the additional cost of computing the viscous
terms and a turbulence model. Finally, Navier–Stokes calculations generally converge much more slowly
than Euler solutions due to discrete stiffness and directional decoupling arising from the highly stretched
boundary layer cells. The computational feasibility of viscous design therefore hinges on the development of
a rapidly convergent Navier–Stokes ﬂowsolver. Pierce and Giles have developed a preconditioned multigrid
method that dramatically improves convergence of viscous calculations by ensuring that all error modes
inside the stretched boundary layer cells are damped efﬁciently [16], [17]. The same acceleration techniques
are applicable to the adjoint calculation, so that the potential payoffs toward reducing the cost of the design
process are substantial.
The ultimate success of an aircraft design depends on the resolution of complex multidisciplinary tradeoffs
between factors such as aerodynamic efﬁciency, structural weight, stability and control, and the volume
required to contain fuel and payload. A design is ﬁnalized only after numerous iterations, cycling between
the disciplines. The development of accurate and efﬁcient methods for aerodynamic shape optimization
represents a worthwhile intermediate step toward the eventual goal of full multidisciplinary optimal design.
2. General Formulation of the Adjoint Approach to Optimal Design
Before embarking on a detailed derivation of the adjoint formulation for optimal design using the Navier–
Stokes equations, it is helpful to summarize the general abstract description of the adjoint approach which
has been thoroughly documented in [3] and [4].
The progress of the design procedure is measured in terms of a cost function I, which could be, for
example, the drag coefﬁcient or the lift to drag ratio. For ﬂow about an airfoil or wing, the aerodynamic
properties which deﬁne the cost function are functions of the ﬂowﬁeld variables (w) and the physical
location of the boundary, which may be represented by the function F, say. Then
I = I(w, F),
and a change in F results in a change
δI =
_
∂I
T
∂w
_
I
δw +
_
∂I
T
∂F
_
II
δF, (1)
in the cost function. Here, the subscripts I and II are used to distinguish the contributions due to the variation
δw in the ﬂow solution from the change associated directly with the modiﬁcation δF in the shape. This
notation is introduced to assist in grouping the numerous terms that arise during the derivation of the full
Navier–Stokes adjoint operator, so that it remains feasible to recognize the basic structure of the approach
as it is sketched in the present section.
Using control theory, the governing equations of the ﬂow ﬁeld are introduced as a constraint in such a
way that the ﬁnal expression for the gradient does not require multiple ﬂow solutions. This corresponds to
eliminating δw from (1).
Suppose that the governing equation Rwhich expresses the dependence of w and F within the ﬂowﬁeld
domain D can be written as
R(w, F) = 0. (2)
Then δw is determined from the equation
δR =
_
∂R
∂w
_
I
δw +
_
∂R
∂F
_
II
δF = 0. (3)
Next, introducing a Lagrange multiplier ψ, we have
216 A. Jameson, L. Martinelli, and N.A. Pierce
δI =
∂I
T
∂w
δw +
∂I
T
∂F
δF −ψ
T
__
∂R
∂w
_
δw +
_
∂R
∂F
_
δF
_
=
_
∂I
T
∂w
−ψ
T
_
∂R
∂w
__
I
δw +
_
∂I
T
∂F
−ψ
T
_
∂R
∂F
__
II
δF. (4)
Choosing ψ to satisfy the adjoint equation
_
∂R
∂w
_
T
ψ =
∂I
∂w
(5)
the ﬁrst term is eliminated, and we ﬁnd that
δI = GδF, (6)
where
G =
∂I
T
∂F
−ψ
T
_
∂R
∂F
_
.
The advantage is that (6) is independent of δw, with the result that the gradient of I with respect to an arbitrary
number of design variables can be determined without the need for additional ﬂowﬁeld evaluations. In the
case that (2) is a partial differential equation, the adjoint equation (5) is also a partial differential equation
and determination of the appropriate boundary conditions requires careful mathematical treatment.
The computational cost of a single design cycle is roughly equivalent to the cost of two ﬂow solutions
since the the adjoint problem has similar complexity. When the number of design variables becomes large,
the computational efﬁciency of the control theory approach over the traditional approach, which requires
direct evaluation of the gradients by individually varying each design variable and recomputing the ﬂow
ﬁeld, becomes compelling.
Once (3) is established, an improvement can be made with a shape change
δF = −λG,
where λ is positive, and small enough that the ﬁrst variation is an accurate estimate of δI. The variation in
the cost function then becomes
δI = −λG
T
G < 0.
After making such a modiﬁcation, the gradient can be recalculated and the process repeated to follow a path
of steepest descent until a minimum is reached. In order to avoid violating constraints, such as a minimum
acceptable wing thickness, the gradient may be projected into an allowable subspace within which the
constraints are satisﬁed. In this way, procedures can be devised which must necessarily converge at least to
a local minimum.
3. The Navier–Stokes Equations
For the derivations that follow, it is convenient to use Cartesian coordinates (x
1
, x
2
, x
3
) and to adopt the
convention of indicial notation where a repeated index “i” implies summation over i = 1 to 3. The three
dimensional Navier–Stokes equations then take the form
∂w
∂t
+
∂f
i
∂x
i
=
∂f
vi
∂x
i
in D, (7)
Optimum Aerodynamic Design Using the Navier–Stokes Equations 217
where the state vector w, inviscid ﬂux vector f and viscous ﬂux vector f
v
are described respectively by
w =
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
ρ
ρu
1
ρu
2
ρu
3
ρE
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
, (8)
f
i
=
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
ρu
i
ρu
i
u
1
+ pδ
i1
ρu
i
u
2
+ pδ
i2
ρu
i
u
3
+ pδ
i3
ρu
i
H
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
, (9)
f
v
i
=
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
0
σ
ij
δ
j1
σ
ij
δ
j2
σ
ij
δ
j3
u
j
σ
ij
+ k
∂T
∂x
i
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
. (10)
In these deﬁnitions, ρ is the density, u
1
, u
2
, u
3
are the Cartesian velocity components, E is the total energy,
and δ
ij
is the Kronecker delta function. The pressure is determined by the equation of state
p = (γ −1)ρ{E −
1
2
(u
i
u
i
)},
and the stagnation enthalpy is given by
H = E +
p
ρ
,
where γ is the ratio of the speciﬁc heats. The viscous stresses may be written as
σ
ij
= µ
_
∂u
i
∂x
j
+
∂u
j
∂x
i
_
+ λδ
ij
∂u
k
∂x
k
, (11)
where µ and λ are the ﬁrst and second coefﬁcients of viscosity. The coefﬁcient of thermal conductivity and
the temperature are computed as
k =
c
p
µ
Pr
, T =
p
Rρ
, (12)
where Pr is the Prandtl number, c
p
is the speciﬁc heat at constant pressure, and R is the gas constant.
For discussion of real applications using a discretization on a body conforming structured mesh, it is also
useful to consider a transformation to the computational coordinates (ξ
1
, ξ
2
, ξ
3
) deﬁned by the metrics
K
ij
=
_
∂x
i
∂ξ
j
_
, J = det(K), K
−1
ij
=
_
∂ξ
i
∂x
j
_
.
The Navier–Stokes equations can then be written in computational space as
∂(Jw)
∂t
+
∂(F
i
−F
vi
)
∂ξ
i
= 0 in D, (13)
where the inviscid and viscous ﬂux contributions are now deﬁned with respect to the computational cell
faces by F
i
= S
ij
f
j
and F
vi
= S
ij
f
vj
, and the quantity S
ij
= JK
−1
ij
is used to represent the projection of
the ξ
i
cell face along the x
j
axis. In obtaining (13) we have made use of the property that
∂S
ij
∂ξ
i
= 0 (14)
which represents the fact that the sum of the face areas over a closed volume is zero, as can be readily
veriﬁed by a direct examination of the metric terms.
218 A. Jameson, L. Martinelli, and N.A. Pierce
4. General Formulation of the Optimal Design Problem for the
Navier–Stokes Equations
Aerodynamic optimization is based on the determination of the effect of shape modiﬁcations on some
performance measure which depends on the ﬂow. For convenience, the coordinates ξ
i
describing the ﬁxed
computational domain are chosen so that each boundary conforms to a constant value of one of these
coordinates. Variations in the shape then result in corresponding variations in the mapping derivatives
deﬁned by K
ij
.
Suppose that the performance is measured by a cost function
I =
_
B
M(w, S) dB
ξ
+
_
D
P(w, S) dD
ξ
,
containing both boundary and ﬁeld contributions where dB
ξ
and dD
ξ
are the surface and volume elements
in the computational domain. In general, Mand P will depend on both the ﬂow variables w and the metrics
S deﬁning the computational space.
The design problem is now treated as a control problem where the boundary shape represents the control
function, which is chosen to minimize I subject to the constraints deﬁned by the ﬂowequations (13). Ashape
change produces a variation in the ﬂow solution δw and the metrics δS which in turn produce a variation in
the cost function
δI =
_
B
δM(w, S) dB
ξ
+
_
D
δP(w, S) dD
ξ
, (15)
with
δM = [M
w
]
I
δw + δM
II
,
δP = [P
w
]
I
δw + δP
II
, (16)
where we continue to use the subscripts I and II to distinguish between the contributions associated with the
variation of the ﬂow solution δw and those associated with the metric variations δS. Thus [M
w
]
I
and [P
w
]
I
represent ∂M/∂w and ∂P/∂w with the metrics ﬁxed, while δM
II
and δP
II
represent the contribution of
the metric variations δS to δMand δP.
In the steady state, the constraint equation (13) speciﬁes the variation of the state vector δw by
∂
∂ξ
i
δ(F
i
−F
vi
) = 0. (17)
Here δF
i
and δF
vi
can also be split into contributions associated with δw and δS using the notation
δF
i
= [F
i
w
]
I
δw + δF
i
II
δF
vi
= [F
vi
w
]
I
δw + δF
vi
II
. (18)
The inviscid contributions are easily evaluated as
[F
i
w
]
I
= S
ij
∂f
j
∂w
, δF
i
II
= δS
ij
f
j
.
The details of the viscous contributions are complicated by the additional level of derivatives in the stress
and heat ﬂux terms and will be derived in Section 6. Multiplying by a costate vector ψ, which will play an
analogous role to the Lagrange multiplier introduced in (4), and integrating over the domain produces
_
D
ψ
T
∂
∂ξ
i
δ(F
i
−F
vi
) = 0. (19)
If ψ is differentiable this may be integrated by parts to give
_
B
n
i
ψ
T
δ(F
i
−F
vi
)dB
ξ
−
_
D
∂ψ
T
∂ξ
i
δ(F
i
−F
vi
) dD
ξ
= 0. (20)
Optimum Aerodynamic Design Using the Navier–Stokes Equations 219
Since the left hand expression equals zero, it may be subtracted from the variation in the cost function (15)
to give
δI =
_
B
[δM−n
i
ψ
T
δ(F
i
−F
vi
)] dB
ξ
+
_
D
[δP +
∂ψ
T
∂ξ
i
δ(F
i
−F
vi
)]dD
ξ
. (21)
Now, since ψ is an arbitrary differentiable function, it may be chosen in such a way that δI no longer
depends explicitly on the variation of the state vector δw. The gradient of the cost function can then be
evaluated directly from the metric variations without having to recompute the variation δw resulting from
the perturbation of each design variable.
Comparing (16) and (18), the variation δw may be eliminated from (21) by equating all ﬁeld terms with
subscript I to produce a differential adjoint system governing ψ
∂ψ
T
∂ξ
i
[F
i
w
−F
vi
w
]
I
+ P
w
= 0 in D. (22)
The corresponding adjoint boundary condition is produced by equating the subscript I boundary terms in
(21) to produce
n
i
ψ
T
[F
i
w
−F
vi
w
]
I
= M
w
on B. (23)
The remaining terms from(21) then yield a simpliﬁed expression for the variation of the cost function which
deﬁnes the gradient
δI =
_
B
{δM
II
−n
i
ψ
T
[δF
i
−δF
vi
]
II
} dB
ξ
+
_
D
_
δP
II
+
∂ψ
T
∂ξ
i
[δF
i
−δF
vi
]
II
_
dD
ξ
. (24)
The details of the formula for the gradient depend on the way in which the boundary shape is parametrized as
a function of the design variables, and the way in which the mesh is deformed as the boundary is modiﬁed.
Using the relationship between the mesh deformation and the surface modiﬁcation, the ﬁeld integral is
reduced to a surface integral by integrating along the coordinate lines emanating from the surface. Thus the
expression for δI is ﬁnally reduced to the form of (6):
δI =
_
B
GδF dB
ξ
,
where F represents the design variables, and G is the gradient, which is a function deﬁned over the boundary
surface.
The boundary conditions satisﬁed by the ﬂow equations restrict the form of the left hand side of the
adjoint boundary condition (23). Consequently, the boundary contribution to the cost function Mcannot be
speciﬁed arbitrarily. Instead, it must be chosen from the class of functions which allow cancellation of all
terms containing δw in the boundary integral of (21). On the other hand, there is no such restriction on the
speciﬁcation of the ﬁeld contribution to the cost function P, since these terms may always be absorbed into
the adjoint ﬁeld equation (22) as source terms.
It is convenient to develop the inviscid and viscous contributions to the adjoint equations separately. Also,
for simplicity, it is assumed that the portion of the boundary that undergoes shape modiﬁcations is restricted
to the coordinate surface ξ
2
= 0. Then (21) and (23) may be simpliﬁed by incorporating the conditions
n
1
= n
3
= 0, n
2
= 1, dB
ξ
= dξ
1
dξ
3
,
so that only the variations δF
2
and δF
v2
need to be considered at the wall boundary.
220 A. Jameson, L. Martinelli, and N.A. Pierce
5. Derivation of the Inviscid Adjoint Terms
The inviscidcontributions have beenpreviouslyderivedin[5] and[18] but are includedhere for completeness.
Taking the transpose of (22), the inviscid adjoint equation may be written as
C
T
i
∂ψ
∂ξ
i
= 0 in D, (25)
where the inviscid Jacobian matrices in the transformed space are given by
C
i
= S
ij
∂f
j
∂w
.
The transformed velocity components have the form
U
i
= S
ij
u
j
,
and the condition that there is no ﬂow through the wall boundary at ξ
2
= 0 is equivalent to
U
2
= 0,
so that
δU
2
= 0
when the boundary shape is modiﬁed. Consequently the variation of the inviscid ﬂux at the boundary
reduces to
δF
2
= δp
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
0
S
21
S
22
S
23
0
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
+ p
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
0
δS
21
δS
22
δS
23
0
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
. (26)
Since δF
2
depends only on the pressure, it is now clear that the performance measure on the boundary
M(w, S) may only be a function of the pressure and metric terms. Otherwise, complete cancellation of the
terms containing δw in the boundary integral would be impossible. For example, arbitrary measures of the
forces and moments may be included in the cost function, since these are functions of the surface pressure.
In order to design a shape which will lead to a desired pressure distribution, a natural choice is to set
I =
1
2
_
B
(p −p
d
)
2
dS,
where p
d
is the desired surface pressure, and the integral is evaluated over the actual surface area. In the
computational domain this is transformed to
I =
1
2
_ _
B
w
(p −p
d
)
2
S
2
 dξ
1
dξ
3
,
where the quantity
S
2
 =
_
S
2j
S
2j
denotes the face area corresponding to a unit element of face area in the computational domain. Now, to
cancel the dependence of the boundary integral on δp, the adjoint boundary condition reduces to
ψ
j
n
j
= p −p
d
, (27)
where n
j
are the components of the surface normal
n
j
=
S
2j
S
2

.
This amounts to a transpiration boundary condition on the costate variables corresponding to the momentum
components. Note that it imposes no restriction on the tangential component of ψ at the boundary.
In the presence of shock waves, neither p nor p
d
are necessarily continuous at the surface. The boundary
condition is then in conﬂict with the assumption that ψ is differentiable. This difﬁculty can be circumvented
by the use of a smoothed boundary condition [18].
Optimum Aerodynamic Design Using the Navier–Stokes Equations 221
6. Derivation of the Viscous Adjoint Terms
In computational coordinates, the viscous terms in the Navier–Stokes equations have the form
∂F
v
i
∂ξ
i
=
∂
∂ξ
i
(S
ij
f
v
j
).
Computing the variation δw resulting from a shape modiﬁcation of the boundary, introducing a costate
vector ψ and integrating by parts following the steps outlined by (17)–(20) produces
_
B
ψ
T
(δS
2j
f
v
j
+ S
2j
δf
v
j
) dB
ξ
−
_
D
∂ψ
T
∂ξ
i
(δS
ij
f
v
j
+ S
ij
δf
v
j
) dD
ξ
,
where the shape modiﬁcation is restricted to the coordinate surface ξ
2
= 0 so that n
1
= n
3
= 0, and n
2
= 1.
Furthermore, it is assumed that the boundary contributions at the far ﬁeld may either be neglected or else
eliminated by a proper choice of boundary conditions as previously shown for the inviscid case [5], [18].
The viscous terms will be derived under the assumption that the viscosity and heat conduction coefﬁcients
µ and k are essentially independent of the ﬂow, and that their variations may be neglected. In the case of
turbulent ﬂow, if the ﬂow variations are found to result in signiﬁcant changes in the turbulent viscosity, it
may eventually be necessary to include its variation in the calculations.
Transformation to Primitive Variables
The derivation of the viscous adjoint terms is simpliﬁed by transforming to the primitive variables
˜ w
T
= (ρ, u
1
, u
2
, u
3
, p)
T
,
because the viscous stresses depend on the velocity derivatives
∂u
i
∂x
j
, while the heat ﬂux can be expressed as
κ
∂
∂x
i
_
p
ρ
_
.
where κ = k/R = γµ/Pr(γ −1). The relationship between the conservative and primitive variations is
deﬁned by the expressions
δw = Mδ ˜ w, δ ˜ w = M
−1
δw
which make use of the transformation matrices M = ∂w −/∂ ˜ w and M
−1
= ∂ ˜ w/∂w. These matrices are
provided in transposed form for future convenience:
M
T
=
_
¸
¸
¸
¸
_
1 u
1
u
2
u
3
u
i
u
i
/2
0 ρ 0 0 ρu
1
0 0 ρ 0 ρu
2
0 0 0 ρ ρu
3
0 0 0 0 1/γ −1
_
¸
¸
¸
¸
_
,
M
−1
T
=
_
¸
¸
¸
¸
_
1 −u
1
/ρ −u
2
/ρ −u
3
/ρ (γ −1)u
i
u
i
/2
0 1/ρ 0 0 −(γ −1)u
1
0 0 1/ρ 0 −(γ −1)u
2
0 0 0 1/ρ −(γ −1)u
3
0 0 0 0 γ −1
_
¸
¸
¸
¸
_
.
The conservative and primitive adjoint operators L and
˜
L corresponding to the variations δw and δ ˜ w are
then related by
_
D
δw
T
Lψ dD
ξ
=
_
D
δ ˜ w
T
˜
Lψ dD
ξ
,
with
˜
L = M
T
L,
222 A. Jameson, L. Martinelli, and N.A. Pierce
so that after determining the primitive adjoint operator by direct evaluation of the viscous portion of (22), the
conservative operator may be obtained by the transformation L = M
−1
T
˜
L. Since the continuity equation
contains no viscous terms, it makes no contribution to the viscous adjoint system. Therefore, the derivation
proceeds by ﬁrst examining the adjoint operators arising from the momentum equations.
Contributions from the Momentum Equations
In order to make use of the summation convention, it is convenient to set ψ
j+1
= ϕ
j
for j = 1, 2, 3. Then the
contribution from the momentum equations is
_
B
ϕ
k
(δS
2j
σ
kj
+ S
2j
δσ
kj
) dB
ξ
−
_
D
∂ϕ
k
∂ξ
i
(δS
ij
σ
kj
+ S
ij
δσ
kj
) dD
ξ
. (28)
The velocity derivatives in the viscous stresses can be expressed as
∂u
i
∂x
j
=
∂u
i
∂ξ
l
∂ξ
l
∂x
j
=
S
lj
J
∂u
i
∂ξ
l
with corresponding variations
δ
∂u
i
∂x
j
=
_
S
lj
J
_
I
∂
∂ξ
l
δu
i
+
_
∂u
i
∂ξ
l
_
II
δ
_
S
lj
J
_
.
The variations in the stresses are then
δσ
kj
=
_
µ
_
S
lj
J
∂
∂ξ
l
δu
k
+
S
lk
J
∂
∂ξ
l
δu
j
_
+ λ
_
δ
jk
S
lm
J
∂
∂ξ
l
δu
m
__
I
+
_
µ
_
δ
_
S
lj
J
_
∂u
k
∂ξ
l
+ δ
_
S
lk
J
_
∂u
j
∂ξ
l
_
+ λ
_
δ
jk
δ
_
S
lm
J
_
∂u
m
∂ξ
l
__
II
.
As before, onlythose terms withsubscript I, whichcontainvariations of the ﬂowvariables, needbe considered
further in deriving the adjoint operator. The ﬁeld contributions that contain δu
i
in equation (28) appear as
−
_
D
∂ϕ
k
∂ξ
i
S
ij
_
µ
_
S
lj
J
∂
∂ξ
l
δu
k
+
S
lk
J
∂
∂ξ
l
δu
j
_
+ λδ
jk
S
lm
J
∂
∂ξ
l
δu
m
_
dD
ξ
.
This may be integrated by parts to yield
_
D
δu
k
∂
∂ξ
l
_
S
lj
S
ij
µ
J
∂ϕ
k
∂ξ
i
_
dD
ξ
+
_
D
δu
j
∂
∂ξ
l
_
S
lk
S
ij
µ
J
∂ϕ
k
∂ξ
i
_
dD
ξ
+
_
D
δu
m
∂
∂ξ
l
_
S
lm
S
ij
λδ
jk
J
∂ϕ
k
∂ξ
i
_
dD
ξ
,
where the boundary integral has been eliminated by noting that δu
i
= 0 on the solid boundary. By exchanging
indices, the ﬁeld integrals may be combined to produce
_
D
δu
k
∂
∂ξ
l
S
lj
_
µ
_
S
ij
J
∂ϕ
k
∂ξ
i
+
S
ik
J
∂ϕ
j
∂ξ
i
_
+ λδ
jk
S
im
J
∂ϕ
m
∂ξ
i
_
dD
ξ
,
which is further simpliﬁed by transforming the inner derivatives back to Cartesian coordinates
_
D
δu
k
∂
∂ξ
l
S
lj
_
µ
_
∂ϕ
k
∂x
j
+
∂ϕ
j
∂x
k
_
+ λδ
jk
∂ϕ
m
∂x
m
_
dD
ξ
. (29)
The boundary contributions that contain δu
i
in (28) may be simpliﬁed using the fact that
∂
∂ξ
l
δu
i
= 0 if l = 1, 3
on the boundary B so that they become
_
B
ϕ
k
S
2j
_
µ
_
S
2j
J
∂
∂ξ
2
δu
k
+
S
2k
J
∂
∂ξ
2
δu
j
_
+ λδ
jk
S
2m
J
∂
∂ξ
2
δu
m
_
dS
x
. (30)
Together, (29) and (30) comprise the ﬁeld and boundary contributions of the momentum equations to the
viscous adjoint operator in primitive variables.
Optimum Aerodynamic Design Using the Navier–Stokes Equations 223
Contributions from the Energy Equation
In order to derive the contribution of the energy equation to the viscous adjoint terms it is convenient to set
ψ
5
= θ, Q
j
= u
i
σ
ij
+ κ
∂
∂x
j
_
p
ρ
_
,
where the temperature has been written in terms of pressure and density using (12). The contribution from
the energy equation can then be written as
_
B
θ
_
δS
2j
Q
j
+ S
2j
δQ
j
_
dB
ξ
−
_
D
∂θ
∂ξ
i
_
δS
ij
Q
j
+ S
ij
δQ
j
_
dD
ξ
. (31)
The ﬁeld contributions that contain δu
i
, δp, and δρ in (31) appear as
−
_
D
∂θ
∂ξ
i
S
ij
δQ
j
dD
ξ
= −
_
D
∂θ
∂ξ
i
S
ij
_
δu
k
σ
kj
+ u
k
δσ
kj
+ κ
S
lj
J
∂
∂ξ
l
_
δp
ρ
−
p
ρ
δρ
ρ
__
dD
ξ
. (32)
The term involving δσ
kj
may be integrated by parts to produce
_
D
δu
k
∂
∂ξ
l
S
lj
_
µ
_
u
k
∂θ
∂x
j
+ u
j
∂θ
∂x
k
_
+ λδ
jk
u
m
∂θ
∂x
m
_
dD
ξ
, (33)
where the conditions u
i
= δu
i
= 0 are used to eliminate the boundary integral on B. Notice that the other
term in (32) that involves δu
k
need not be integrated by parts and is merely carried on as
−
_
D
δu
k
σ
kj
S
ij
∂θ
∂ξ
i
dD
ξ
. (34)
The terms in expression (32) that involve δp and δρ may also be integrated by parts to produce both a
ﬁeld and a boundary integral. The ﬁeld integral becomes
_
D
_
δp
ρ
−
p
ρ
δρ
ρ
_
∂
∂ξ
l
_
S
lj
S
ij
κ
J
∂θ
∂ξ
i
_
dD
ξ
,
which may be simpliﬁed by transforming the inner derivative to Cartesian coordinates
_
D
_
δp
ρ
−
p
ρ
δρ
ρ
_
∂
∂ξ
l
_
S
lj
κ
∂θ
∂x
j
_
dD
ξ
. (35)
The boundary integral becomes
_
B
κ
_
δp
ρ
−
p
ρ
δρ
ρ
_
S
2j
S
ij
J
∂θ
∂ξ
i
dB
ξ
. (36)
This can be simpliﬁed by transforming the inner derivative to Cartesian coordinates
_
B
κ
_
δp
ρ
−
p
ρ
δρ
ρ
_
S
2j
J
∂θ
∂x
j
dB
ξ
, (37)
and identifying the normal derivative at the wall
∂
∂n
= S
2j
∂
∂x
j
, (38)
and the variation in temperature
δT =
1
R
_
δp
ρ
−
p
ρ
δρ
ρ
_
,
to produce the boundary contribution
_
B
kδT
∂θ
∂n
dB
ξ
. (39)
This term vanishes if T is constant on the wall but persists if the wall is adiabatic.
224 A. Jameson, L. Martinelli, and N.A. Pierce
There is also a boundary contribution left over from the ﬁrst integration by parts (31) which has the form
_
B
θδ(S
2j
Q
j
) dB
ξ
, (40)
where
Q
j
= k
∂T
∂x
j
,
since u
i
= 0. Notice that for future convenience in discussing the adjoint boundary conditions resulting from
the energy equation, both the δw and δS terms corresponding to subscript classes I and II are considered
simultaneously. If the wall is adiabatic
∂T
∂n
= 0,
so that, using (38),
δ(S
2j
Q
j
) = 0,
and both the δw and δS boundary contributions vanish.
On the other hand, if T is constant, ∂T/∂ξ
l
= 0 for l = 1, 3, so that
Q
j
= k
∂T
∂x
j
= k
_
S
l
j
J
∂T
∂ξ
l
_
= k
_
S
2j
J
∂T
∂ξ
2
_
.
Thus, the boundary integral (40) becomes
_
B
kθ
_
S
2
2j
J
∂
∂ξ
2
δT + δ
_
S
2
2j
J
_
∂T
∂ξ
2
_
dB
ξ
. (41)
Therefore, for constant T, the ﬁrst term corresponding to variations in the ﬂow ﬁeld contributes to the
adjoint boundary operator and the second set of terms corresponding to metric variations contribute to the
cost function gradient.
Altogether, the contributions from the energy equation to the viscous adjoint operator are the three ﬁeld
terms (33), (34) and (35), and either of two boundary contributions (39) or (41), depending on whether the
wall is adiabatic or has constant temperature.
7. The Viscous Adjoint Field Operator
Collecting together the contributions fromthe momentumand energy equations, the viscous adjoint operator
in primitive variables can be expressed as
(
˜
Lψ)
1
= −
p
ρ
2
∂
∂ξ
l
_
S
lj
κ
∂θ
∂x
j
_
(
˜
Lψ)
i+1
=
∂
∂ξ
l
_
S
lj
_
µ
_
∂ϕ
i
∂x
j
+
∂ϕ
j
∂x
i
_
+ λδ
ij
∂ϕ
k
∂x
k
__
i = 1, 2, 3
+
∂
∂ξ
l
_
S
lj
_
µ
_
u
i
∂θ
∂x
j
+ u
j
∂θ
∂x
i
_
+ λδ
ij
u
k
∂θ
∂x
k
__
−σ
ij
S
lj
∂θ
∂ξ
l
,
(
˜
Lψ)
5
= ρ
∂
∂ξ
l
_
S
lj
κ
∂θ
∂x
j
_
.
The conservative viscous adjoint operator may now be obtained by the transformation
L = M
−1
T
˜
L.
Optimum Aerodynamic Design Using the Navier–Stokes Equations 225
8. Viscous Adjoint Boundary Conditions
It was recognized in Section 4 that the boundary conditions satisﬁed by the ﬂow equations restrict the form
of the performance measure that may be chosen for the cost function. There must be a direct correspondence
between the ﬂowvariables for which variations appear in the variation of the cost function, and those variables
for which variations appear in the boundary terms arising during the derivation of the adjoint ﬁeld equations.
Otherwise it would be impossible to eliminate the dependence of δI on δw through proper speciﬁcation of
the adjoint boundary condition. As in the derivation of the ﬁeld equations, it proves convenient to consider
the contributions from the momentum equations and the energy equation separately.
Boundary Conditions Arising from the Momentum Equations
The boundary term that arises from the momentum equations including both the δw and δS components
(28) takes the form
_
B
ϕ
k
δ(S
2j
σ
kj
) dB
ξ
.
Replacing the metric term with the corresponding local face area S
2
and unit normal n
j
deﬁned by
S
2
 =
_
S
2j
S
2j
, n
j
=
S
2j
S
2

then leads to
_
B
ϕ
k
δ(S
2
n
j
σ
kj
) dB
ξ
.
Deﬁning the components of the surface stress as
τ
k
= n
j
σ
kj
and the physical surface element
dS = S
2
 dB
ξ
,
the integral may then be split into two components
_
B
ϕ
k
τ
k
δS
2
 dB
ξ
+
_
B
ϕ
k
S
2
δτ
k
dS, (42)
where only the second term contains variations in the ﬂow variables and must consequently cancel the δw
terms arising in the cost function. The ﬁrst term will appear in the expression for the gradient.
Ageneral expression for the cost function that allows cancellation with terms containing δτ
k
has the form
I =
_
B
N(τ) dS, (43)
corresponding to a variation
δI =
_
B
∂N
∂τ
k
δτ
k
dS,
for which cancellation is achieved by the adjoint boundary condition
ϕ
k
=
∂N
∂τ
k
.
Natural choices for N arise from force optimization and as measures of the deviation of the surface stresses
from desired target values.
For viscous force optimization, the cost function should measure friction drag. The friction force in the
x
i
direction is
CD
fi
=
_
B
σ
ij
dS
j
=
_
B
S
2j
σ
ij
dB
ξ
226 A. Jameson, L. Martinelli, and N.A. Pierce
so that the force in a direction with cosines n
i
has the form
C
nf
=
_
B
n
i
S
2j
σ
ij
dB
ξ
.
Expressed in terms of the surface stress τ
i
, this corresponds to
C
nf
=
_
B
n
i
τ
i
dS,
so that basing the cost function (43) on this quantity gives
N = n
i
τ
i
.
Cancellation with the ﬂow variation terms in (42) therefore mandates the adjoint boundary condition
ϕ
k
= n
k
.
Note that this choice of boundary condition also eliminates the ﬁrst term in (42) so that it need not be
included in the gradient calculation.
In the inverse design case, where the cost function is intended to measure the deviation of the surface
stresses from some desired target values, a suitable deﬁnition is
N(τ) =
1
2
a
lk
(τ
l
−τ
dl
)(τ
k
−τ
dk
),
where τ
d
is the desired surface stress, including the contribution of the pressure, and the coefﬁcients a
lk
deﬁne a weighting matrix. For cancellation
ϕ
k
δτ
k
= a
lk
(τ
l
−τ
dl
)δτ
k
.
This is satisﬁed by the boundary condition
ϕ
k
= a
lk
(τ
l
−τ
dl
). (44)
Assuming arbitrary variations in δτ
k
, this condition is also necessary.
In order to control the surface pressure and normal stress one can measure the difference
n
j
{σ
kj
+ δ
kj
(p −p
d
)},
where p
d
is the desired pressure. The normal component is then
τ
n
= n
k
n
j
σ
kj
+ p −p
d
,
so that the measure becomes
N(τ) =
1
2
τ
2
n
=
1
2
n
l
n
m
n
k
n
j
{σ
lm
+ δ
lm
(p −p
d
)}{σ
kj
+ δ
kj
(p −p
d
)}.
This corresponds to setting
a
lk
= n
l
n
k
in (44). Deﬁning the viscous normal stress as
τ
vn
= n
k
n
j
σ
kj
,
the measure can be expanded as
N(τ) =
1
2
n
l
n
m
n
k
n
j
σ
lm
σ
kj
+
1
2
(n
k
n
j
σ
kj
+ n
l
n
m
σ
lm
)(p −p
d
) +
1
2
(p −p
d
)
2
=
1
2
τ
2
vn
+ τ
vn
(p −p
d
) +
1
2
(p −p
d
)
2
.
Optimum Aerodynamic Design Using the Navier–Stokes Equations 227
For cancellation of the boundary terms
ϕ
k
(n
j
δσ
kj
+ n
k
δp) = {n
l
n
m
σ
lm
+ n
2
l
(p −p
d
)}n
k
(n
j
δσ
kj
+ n
k
δ
p
)
leading to the boundary condition
ϕ
k
= n
k
(τ
vn
+ p −p
d
).
In the case of high Reynolds number, this is well approximated by the equations
ϕ
k
= n
k
(p −p
d
), (45)
which should be compared with the single scalar equation derived for the inviscid boundary condition (27).
In the case of an inviscid ﬂow, choosing
N(τ) =
1
2
(p −p
d
)
2
requires
ϕ
k
n
k
δp = (p −p
d
)n
2
k
δp = (p −p
d
)δp
which is satisﬁed by (45), but which represents an overspeciﬁcation of the boundary condition since only
the single condition (27) need be speciﬁed to ensure cancellation.
Boundary Conditions Arising from the Energy Equation
The form of the boundary terms arising from the energy equation depends on the choice of temperature
boundary condition at the wall. For the adiabatic case, the boundary contribution is (39)
_
B
kδT
∂θ
∂n
dB
ξ
,
while for the constant temperature case the boundarytermis (41). One possibilityis tointroduce a contribution
into the cost function which is dependent T or ∂T/∂n so that the appropriate cancellation would occur.
Since there is little physical intuition to guide the choice of such a cost function for aerodynamic design, a
more natural solution is to set
θ = 0
in the constant temperature case or
∂θ
∂n
= 0
in the adiabatic case. Note that in the constant temperature case, this choice of θ on the boundary would also
eliminate the boundary metric variation terms in (40).
9. Implementation of Navier–Stokes Design
The design procedures can be summarized as follows:
1. Solve the ﬂow equations for ρ, u
1
, u
2
,u
3
, p.
2. Smooth the cost function, if necessary.
3. Solve the adjoint equations for ψ subject to appropriate boundary conditions.
4. Evaluate G .
5. Project G into an allowable subspace that satisﬁes any geometric constraints.
6. Update the shape based on the direction of steepest descent.
7. Return to 1.
Practical implementation of the viscous design method relies heavily upon fast and accurate solvers
for both the state (w) and costate (ψ) systems. This work employs a wellvalidated Navier–Stokes solver
developed by two of the authors [19].
228 A. Jameson, L. Martinelli, and N.A. Pierce
Discretization
Both the ﬂow and the adjoint equations are discretized using a semidiscrete cellcentered ﬁnite volume
scheme. The convective ﬂuxes across cell interfaces are represented by simple arithmetic averages of the
ﬂuxes computed using values from the cells on either side of the face, augmented by artiﬁcial diffusive
terms to prevent numerical oscillations in the vicinity of shock waves. Continuing to use the summation
convention for repeated indices, the numerical convective ﬂux across the interface between cells A and B in
a three dimensional mesh has the form
h
AB
=
1
2
S
AB
j
(f
A
j
+ f
B
j
) −d
AB
,
where S
AB
j
is the component of the face area in the jth Cartesian coordinate direction, (f
A
j
) and (f
B
j
)
denote the ﬂux f
j
as deﬁned by (12) and d
AB
is the diffusive term. Variations of the computer program
provide options for alternate constructions of the diffusive ﬂux.
The simplest option implements the JamesonSchmidtTurkel scheme [20], [21], using scalar diffusive
terms of the form
d
AB
=
(2)
∆w −
(4)
(∆w
+
−2∆w + ∆w
−
),
where
∆w = w
B
−w
A
and ∆w
+
and ∆w
−
are the same differences across the adjacent cell interfaces behind cell A and beyond
cell B in the AB direction. By making the coefﬁcient
(2)
depend on a switch proportional to the undivided
second difference of a ﬂow quantity such as the pressure or entropy, the diffusive ﬂux becomes a third order
quantity, proportional to the cube of the mesh width in regions where the solution is smooth. Oscillations
are suppressed near a shock wave because
(2)
becomes of order unity, while
(4)
is reduced to zero by the
same switch. For a scalar conservation law, it is shown in [21] that
(2)
and
(4)
can be constructed to make
the scheme satisfy the local extremumdiminishing (LED) principle that local maxima cannot increase while
local minima cannot decrease.
The second option applies the same construction to local characteristic variables. There are derived from
the eigenvectors of the Jacobian matrix A
AB
which exactly satisﬁes the relation
A
AB
(w
B
−w
A
) = S
AB
j
(f
B
j
−f
A
j
).
This corresponds to the deﬁnition of Roe [22]. The resulting scheme is LED in the characteristic variables.
The third option implements the HCUSP scheme proposed by Jameson [23] which combines differences
f
B
−f
A
and w
B
−w
A
in a manner such that stationary shock waves can be captured with a single interior
point in the discrete solution. This scheme minimizes the numerical diffusion as the velocity approaches
zero in the boundary layer, and has therefore been preferred for viscous calculations in this work.
Similar artiﬁcial diffusive terms are introduced in the discretization of the adjoint equation, but with the
opposite sign because the wave directions are reversed in the adjoint equation. Satisfactory results have been
obtained using scalar diffusion in the adjoint equation while characteristic or HCUSP constructions are
used in the ﬂow solution.
The discretization of the viscous terms of the Navier–Stokes equations requires the evaluation of the
velocity derivatives ∂
u
i
/∂
x
j
in order to calculate the viscous stress tensor σ
ij
deﬁned in (11). These are most
conveniently evaluated at the cell vertices of the primary mesh by introducing a dual mesh which connects
the cell centers of the primary mesh, as depicted in Figure 1. According to the Gauss formula for a control
volume V with boundary S
_
V
∂v
i
∂x
j
dv =
_
S
u
i
n
j
dS,
where n
j
is the outward normal. Applied to the dual cells this yields the estimate
∂v
i
∂x
j
=
1
vol
faces
¯ u
i
n
j
S,
where S is the area of a face, and ¯ u
i
is an estimate of the average of u
i
over that face. In order to determine
the viscous ﬂux balance of each primary cell, the viscous ﬂux across each of its faces is then calculated from
Optimum Aerodynamic Design Using the Navier–Stokes Equations 229
i j
σ
dual cell
Figure 1. Cellcentered scheme. σ
ij
evaluated at vertices of the primary mesh.
the average of the viscous stress tensor at the four vertices connected by that face. This leads to a compact
scheme with a stencil connecting each cell to its 26 nearest neighbors.
The semidiscrete schemes for both the ﬂow and the adjoint equations are both advanced to steady state
by a multistage time stepping scheme. This is a generalized Runge–Kutta scheme in which the convective
and diffusive terms are treated differently to enlarge the stability region [21], [24]. Convergence to a steady
state is accelerated by residual averaging and a multigrid procedure [25]. Convergence is further accelerated
by the use of locally varying time steps (which may be regarded as a scalar preconditioner) or the matrix
preconditioner method developed by Pierce and Giles [16], [17].
Optimization
For inverse design the lift is ﬁxed by the target pressure. In drag minimization it is also appropriate to ﬁx
the lift coefﬁcient, because the induced drag is a major fraction of the total drag, and this could be reduced
simply by reducing the lift. Therefore the angle of attack is adjusted during the ﬂow solution to force a
speciﬁed lift coefﬁcient to be attained.
The search procedure used in this work is a simple descent method in which small steps are taken in the
negative gradient direction. Let F represent the design variable, and G the gradient. Then the iteration
δF = −λG
can be regarded as simulating the time dependent process
dF
dt
= −G,
where λ is the time step ∆t. Let A be the Hessian matrix with element
A
ij
=
∂G
i
∂F
j
=
∂
2
I
∂F
i
∂F
j
.
Suppose that a locally minimum value of the cost function I
∗
= I(F
∗
) is attained when F = F
∗
. Then the
gradient G
∗
= G(F
∗
) must be zero, while the Hessian matrix A
∗
= A(F
∗
) must be positive deﬁnite. Since
G
∗
is zero, the cost function can be expanded as a Taylor series in the neighborhood of F
∗
with the form
I(F) = I
∗
+
1
2
(F −F
∗
)A(F −F
∗
) + · · · .
Correspondingly,
G(F) = A(F −F
∗
) + · · · .
As F approaches F
∗
, the leading terms become dominant. Then, setting
ˆ
F = (F −F
∗
), the search process
230 A. Jameson, L. Martinelli, and N.A. Pierce
approximates
d
ˆ
F
dt
= −A
∗
ˆ
F.
Also, since A
∗
is positive deﬁnite it can be expanded as
A
∗
= RMR
T
,
where M is a diagonal matrix containing the eigenvalues of A
∗
, and
RR
T
= R
T
R = I.
Setting
v = R
T
ˆ
F,
the search process can be represented as
dv
dt
= −Mv.
The stability region for the simple forward Euler stepping scheme is a unit circle centered at −1 on the
negative real axis. Thus for stability we must choose
µ
max
∆t = µ
max
λ < 2,
while the asymptotic decay rate, given by the smallest eigenvalue, is proportional to
e
−µ
min
t
.
In order to improve the rate of convergence, one can set
δF = −λPG,
where P is a preconditioner for the search. An ideal choice is P = A
∗−1
, so that the corresponding time
dependent process reduces to
d
ˆ
F
dt
= −
ˆ
F,
for which all the eigenvalues are equal to unity, and
ˆ
F is reduced to zero in one time step by the choice
∆t = 1. QuasiNewton methods estimate A
∗
from the change in the gradient during the search process. This
requires accurate estimates of the gradient at each time step. In order to obtain these, both the ﬂow solution
and the adjoint equation must be fully converged. Most quasiNewton methods also require a line search in
each search direction, for which the ﬂow equations and cost function must be accurately evaluated several
times. They have proven quite robust for aerodynamic optimization [7].
An alternative approach which has also proved successful in our previous work [18], and is used here, is
to smooth the gradient and to replace G by its smoothed value
¯
G in the descent process. This acts both as
a preconditioner and ensures that each new shape in the optimization sequence remains smooth. To apply
smoothing in the ξ
1
direction, for example, the smoothed gradient
¯
G may be calculated from a discrete
approximation to
¯
G −
∂
∂ξ
1
∂
∂ξ
1
¯
G = G,
where is the smoothing parameter. If δF = −λ
¯
G is set, then, assuming the modiﬁcation is applied on the
surface ξ
2
= constant, the ﬁrst order change in the cost function is
δI = −
_ _
GδF dξ
1
dξ
3
= −λ
_ _ _
¯
G −
∂
∂ξ
1
∂
¯
G
∂ξ
1
_
¯
G dξ
1
dξ
3
= −λ
_ _ _
¯
G
2
+
_
∂
¯
G
∂ξ
1
_
2
_
dξ
1
dξ
3
< 0,
Optimum Aerodynamic Design Using the Navier–Stokes Equations 231
assuring an improvement if λ is sufﬁciently small and positive, unless the process has already reached a
stationary point at which G = 0.
It turns out that this approach is tolerant to the use of approximate values of the gradient, so that neither the
ﬂow solution nor the adjoint solution need be fully converged before making a shape change. This results in
very large savings in the computational cost. For inviscid optimization it is necessary to use only 15 multigrid
cycles for the ﬂow solution and the adjoint solution in each design iteration. For viscous optimization, about
100 multigrid cycles are needed. This is partly because convergence of the lift coefﬁcient is much slower,
so about 20 iterations must be made before each adjustment of the angle of attack to force the target lift
coefﬁcient. The new preconditioner for the ﬂow and adjoint calculations allows the number of iterations to
be substantially reduced in both the ﬂow and the adjoint simulation.
The numerical tests so far have focused on the viscous design of wings for optimum cruise, for which
the ﬂow remains attached, and the main viscous effect is due to the displacement thickness of the boundary
layer. While some tests have been made with the viscous adjoint terms included, it has been found that
the optimization process converges when the viscous terms are omitted from the adjoint system. This may
reﬂect the tolerance of the search process to inexact gradients.
10. Results
Preconditioned Inverse Design
The ﬁrst demonstration is an application of the preconditioning technique for inverse design with the Euler
equations. The ONERA M6 (Figure 2(b)) wing is recovered for a lifting case starting from a wing with an
NACA0012 section (Figure 2(a)) and using the ONERA M6 pressure distributions computed at α = 3.0 and
M = 0.84 as the target (Figure 3). Thus, a symmetric wing section is to be recovered from an asymmetric
pressure distribution. The calculations were performed on a 192 × 32 × 48 mesh with 294,912 cells. The
Figure 2. Redesign of the Onera M6 wing. 100 design cycles in inverse mode. (a) M = 0.84, C
L
= 0.3000, C
D
= 0.0205, α = 2.935°.
(b) M = 0.84, C
L
= 0.2967, C
D
= 0.0141, α = 2.935°.
232 A. Jameson, L. Martinelli, and N.A. Pierce
1
.2
0
.8
0
.4
0
.0
0
.4
0
.8
0
1
.2
0
1
.6
0
2
.0
0
C
p
+
+
+
+
+
+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+
+
+
+
+
+
+
+
+
+ + +
+
+
+
+
+
+
+
+
+
+
+++++
+
+
+
+
++
+
++++++++++++++++++++++++++++
+
+
+++ + + +
+
+
+
+
+
+o
o
o
o
o
o
o
o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
ooooo
o
o
o
o
o
o
oooooooooooooooooooooooooooo o
o
o
o o o o o
o
o
o
o
o
o
o
1
.2
0
.8
0
.4
0
.0
0
.4
0
.8
0
1
.2
0
1
.6
0
2
.0
0
C
p
+
+
+
+
+
+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+
+
+
+
+
+
+
+
+
+ + +
+
+
+
+
+
+
+
+
+
+
+++++++
+
+
+
+
+
+
+
+++++++++++++++++++++++
+
+
++++++ + + + +
+
+
+
+
+o
o
o
o
o
o
o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
ooooo
o
o
ooo
o
o
o
o
oooooooooooooooooooooo
o
o
o o o o o o o o o
o
o
o
o
o
o
1
.2
0
.8
0
.4
0
.0
0
.4
0
.8
0
1
.2
0
1
.6
0
2
.0
0
C
p
+
+
+
+
+
+ + + + + + + +
+
+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+
+
+
+
+
+
+
+
+
+ +
+
+
+
+
+
+
+
+
+
+
+
++++++++++++++++++
+
+
+
+++
+
+++++++++
+
+
+++++++++ + + + +
+
+
+
+
+o
o
o
o
o
o
o o o o o o o
o
o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
oooooo
ooooooooooo
o
o
o
oooo
o
oooooooo
o
o
ooo o o o o o o o o o
o
o
o
o
o
o
1
.2
0
.8
0
.4
0
.0
0
.4
0
.8
0
1
.2
0
1
.6
0
2
.0
0
C
p
+
+
+
+
+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+
+
+
+
+
+
+
+
+
+ +
+
+
+
+
+
+
+
+
+
+
+
+
+++++++++++++++++++++++
+
+
+
++++++++++++++++++ + + +
+
+
+
+
+
+o
o
o
o
o
o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
oooooooooooooooooooooooo
o
o
o
ooooooooooo o o o o o o o o o
o
o
o
o
o
o
(a) (b)
(c) (d)
Figure 3. Target (◦) and computed (+) pressure distributions of redesigned Onera M6 wing. M = 0.84, C
L
= 0.2967, C
D
= 0.0141,
α = 2.935
◦
.
mesh had a CH topology with the Clines wrapping around the wing leading edge. Each design cycle
required 3 multigrid cycles for the ﬂow solver using characteristicbased matrix dissipation with a matrix
preconditioner and 12 multigrid cycles for the adjoint solver using scalar dissipation and a variable local
time step (scalar preconditioner). Compared with a test in which the 3 multigrid cycles using the matrix
preconditioner were replaced by 15 multigrid cycles using a standard scalar preconditioner, and 15 cycles
were used in the adjoint solution, each design cycle required about threeeights as much computer time,
while the number of design cycles required to reach the same level of error also fell from 100 to about 50.
Use of the matrix preconditioner therefore reduced the total CPU time on an IBM 590 workstation from
97,683 s (∼27 hours) to 18,222 s (∼5 hours) for roughly equivalent accuracy.
Viscous Design
Due to the high computational cost of viscous design, a twostage design strategy is adopted. In the ﬁrst stage
a design calculation is performed with the Euler equations to minimize the drag at a given lift coefﬁcient
by modifying the wing sections with a ﬁxed planform. In the second stage the pressure distribution of the
Euler solution is used as the target pressure distribution for inverse design with the Navier–Stokes equations.
Comparatively small modiﬁcations are required in the second stage, so that it can be accomplished with a
small number of design cycles.
In order to test this strategy it was used for the redesign of a wing representative of widebody transport
aircraft. The results are shown in Figures 4 and 5. The design point was taken as a lift coefﬁcient of 0.55
at a Mach number of 0.83. Figure 4 illustrates the Euler redesign, which was performed on a mesh with
Optimum Aerodynamic Design Using the Navier–Stokes Equations 233
Figure 4. Redesign of the wing of a wide transport aircraft. Stage
1: inviscid design. Sixty design cycles in drag reduction mode with
forced lift. (a) M = 0.83, C
L
= 0.5498, C
D
= 0.0196, α = 2.410
◦
.
(b) M = 0.83, C
L
= 0.5500, C
d
= 0.0181, α = 1.959
◦
.
Figure 5. Redesign of the wing of a wide transport aircraft. Stage
2: viscous redesign. Ten design cycles in inverse mode. (a) M =
0.83, C
L
= 0.5506, C
D
= 0.0199, α = 2.317
◦
. (b) M = 0.83,
C
L
= 0.5508, C
D
= 0.0194, α = 2.355
◦
.
192 ×32 ×48 cells, displaying both the geometry and the upper surface pressure distribution, with negative
Cp upwards. The initial wing shows a moderately strong shock wave across most of the top surface, as can
be seen in Figure 4(a). Sixty design cycles were needed to produce the shock free wing shown in Figure 4(b),
with an indicated drag reduction of 15 counts from 0.0196 to 0.0181. Figure 5 shows the viscous redesign at
a Reynolds number of 12 million. This was performed on a mesh with 192 ×64×48 cells, with 32 intervals
normal to the wing concentrated inside the boundary layer region. In Figure 5(a) it can be seen that the
Euler design produces a weak shock due to the displacement effects of the boundary layer. Ten design cycles
were needed to recover the shock free wing shown in Figure 5(b). It is interesting that the wing section
modiﬁcations between the initial wing of Figure 4(a) and the ﬁnal wing of Figure 5(b) are remarkably small.
These results were sufﬁciently promising that it was decided by McDonnell Douglas to evaluate the
method for industrial use, and it was used to support design studies for the MDXX project. The results of
this experience are discussed in [26]. It rapidly became apparent that the fuselage effects are too large to be
ignored. In viscous design it was also found that there were discrepancies between the results of the design
calculations, which were initially performed on a relatively coarse grid with 192 × 64 × 48 cells, and the
results of subsequent analysis calculations performed on ﬁner meshes to verify the design.
In order to allow the use of ﬁner meshes with overnight turnaround, the code was therefore modiﬁed for
parallel computation. Using the parallel implementation, viscous design calculations have been performed
on meshes with 1.8 million mesh points. Starting from a preliminary inviscid design, 20 design cycles are
usually sufﬁcient for a viscous redesign in inverse mode, with the smoothed inviscid results providing the
target pressure. Such a calculation can be completed in about 7
1
2
hours using 48 processors of an IBM SP2.
As an illustration of the results that could be obtained, Figures 6–10 showa wingbody design with sweep
back of about 38 degrees at the 1/4 chord. Starting fromthe result of an Euler design, the viscous optimization
234 A. Jameson, L. Martinelli, and N.A. Pierce
SYMBOL
SOURCE
SYN107P DESIGN40
ALPHA
2.094
CL
0.610
CD
0.01126
COMPARISON OF CHORDWISE PRESSURE DISTRIBUTIONS
MPX5X WINGBODY
REN = 101.00 , MACH = 0.860
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
Solution 1
UpperSurface Isobars
( Contours at 0.05 Cp )
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
9.6% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
27.2% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
44.3% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
60.8% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
79.1% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
95.6% Span
Figure 6. Pressure distribution of the MPX5X at its design point.
SYMBOL
SOURCE
SYN107P DESIGN20
SYN107P DESIGN10
SYN107P DESIGN 0
ALPHA
2.113
2.153
2.251
CD
0.01127
0.01125
0.01131
COMPARISON OF CHORDWISE PRESSURE DISTRIBUTIONS
MPX5X WINGBODY
REN = 101.00 , MACH = 0.860 , CL = 0.610
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
Solution 1
UpperSurface Isobars
( Contours at 0.05 Cp )
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
9.6% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
27.2% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
44.3% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
60.8% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
79.1% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
95.6% Span
Figure 7. Optimization Sequence in the design of the MPX5X.
Optimum Aerodynamic Design Using the Navier–Stokes Equations 235
SYMBOL
SOURCE
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
MACH
0.860
0.855
0.850
0.845
0.840
ALPHA
2.098
2.173
2.231
2.280
2.326
CD
0.01121
0.01140
0.01141
0.01130
0.01114
COMPARISON OF CHORDWISE PRESSURE DISTRIBUTIONS
MPX5X WINGBODY
REN = 101.00 , CL = 0.610
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
Solution 1
UpperSurface Isobars
( Contours at 0.05 Cp )
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
9.6% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
27.2% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
44.3% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
60.8% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
79.1% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
95.6% Span
Figure 8. Off design performance of the MPX5X below the design point.
SYMBOL
SOURCE
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
MACH
0.860
0.865
0.870
0.880
ALPHA
2.098
2.032
1.962
1.819
CD
0.01121
0.01156
0.01219
0.01423
COMPARISON OF CHORDWISE PRESSURE DISTRIBUTIONS
MPX5X WINGBODY
REN = 101.00 , CL = 0.610
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
Solution 1
UpperSurface Isobars
( Contours at 0.05 Cp )
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
9.6% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
27.2% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
44.3% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
60.8% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
79.1% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
95.6% Span
Figure 9. Off design performance of the MPX5X above the design point.
236 A. Jameson, L. Martinelli, and N.A. Pierce
SYMBOL
SOURCE
MPX5X DESIGN 40
MPX5X DESIGN 40
MPX5X DESIGN 40
ALPHA
2.098
1.814
2.380
CL
0.610
0.558
0.661
CD
0.01121
0.01032
0.01314
COMPARISON OF CHORDWISE PRESSURE DISTRIBUTIONS
MPX5X WINGBODY
REN = 101.00 , MACH = 0.860
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
COMPPLOT
JCV 1.13
Solution 1
UpperSurface Isobars
( Contours at 0.05 Cp )
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
9.6% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
27.2% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
44.3% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
60.8% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
79.1% Span
0.2 0.4 0.6 0.8 1.0
1.0
0.8
0.5
0.2
0.0
0.2
0.5
0.8
1.0
C
p
X / C
95.6% Span
Figure 10. Comparison of the MPX5X at its design point and at lower and higher lift.
produced an essentially shockfree wing at a cruise design point of Mach 0.86, with a lift coefﬁcient of 0.6
for the wing body combination at a Reynolds number of 101 million based on the root chord. Figure 6 shows
the design point, while the evolution of the design is shown in Figure 7, using Vassberg’s COMPPLOT
software. In this case the pressure contours are for the ﬁnal design. This wing is quite thick, with a thickness
to chord ratio of more than 14 % at the root and 9 % at the tip. The design offers excellent performance at
the nominal cruise point. Figures 8 and 9 show the results of a Mach number sweep to determine the drag
rise. It can be seen that a double shock pattern forms below the design point, while there is actually a slight
increase in the drag coefﬁcient of about 1
1
2
counts at Mach 0.85. Finally, Figure 10 shows a comparison of
the pressure distribution at the design point with those at alternate cruise points with lower and higher lift.
The tendency to produce double shocks below the design point is typical of supercritical wings. This wing
has a low drag coefﬁcient, however, over a wide range of conditions.
11. Conclusions
We have developed a threedimensional control theory based design method for the Navier–Stokes equations
and applied it successfully to the design of wings in transonic ﬂow. The method represents an extension of our
previous workondesignwiththe potential ﬂowandEuler equations. The newmethodcombines the versatility
of numerical optimization methods with the efﬁciency of inverse design. The geometry is modiﬁed by a grid
perturbation technique which is applicable to arbitrary conﬁgurations. The combination of computational
efﬁciency with geometric ﬂexibility provides a powerful tool, with the ﬁnal goal being to create practical
aerodynamic shape design methods for complete aircraft conﬁgurations. Such an accomplishment would
represent the culmination of the line of research initiated by Lighthill with his original work on the inverse
problem [1].
Optimum Aerodynamic Design Using the Navier–Stokes Equations 237
References
[1] M.J. Lighthill. A new method of two dimensional aerodynamic design. R & M 1111, Aeronautical Research Council, 1945.
[2] J.L. Lions. Optimal Control of Systems Governed by Partial Differential Equations. SpringerVerlag, New York, 1971. Translated
by S.K. Mitter.
[3] A. Jameson. Aerodynamic design via control theory. J. Sci. Comput., 3:233–260, 1988.
[4] A. Jameson. Optimum aerodynamic design using CFD and control theory. AIAA Paper 951729CP, 1995.
[5] A. Jameson. Automatic design of transonic airfoils to reduce the shock induced pressure drag. In Proceedings of the 31st Israel
Annual Conference on Aviation and Aeronautics, Tel Aviv, pages 5–17, February 1990.
[6] J. Reuther and A. Jameson. Control based airfoil design using the Euler equations. AIAA paper 944272CP, 1994.
[7] J. Reuther and A. Jameson. Aerodynamic shape optimization of wing and wing–body conﬁgurations using control theory. AIAA
Paper 950123, AIAA 33rd Aerospace Sciences Meeting, Reno, NV, January 1995.
[8] J. Reuther, A. Jameson, J. Farmer, L. Martinelli, and D. Saunders. Aerodynamic shape optimization of complex aircraft conﬁgu
rations via an adjoint method. AIAA paper 960094, AIAA 34th Aerospace Sciences Meeting, Reno, NV, January 1996.
[9] O. Pironneau. Optimal Shape Design for Elliptic Systems. SpringerVerlag, New York, 1984.
[10] O. Pironneau. Optimal shape design for aerodynamics. In AGARD REPORT 803, 1994.
[11] S. Ta’asan, G. Kuruvila, and M. D. Salas. Aerodynamic design and optimization in one shot. AIAA paper 92005, 30th Aerospace
Sciences Meeting and Exibit, Reno, NV, January 1992.
[12] O. Baysal and M. E. Eleshaky. Aerodynamic design optimization using sensitivity analysis and computational ﬂuid dynamics.
AIAA J., 30(3):718–725, 1992.
[13] H. Cabuk, C.H. Shung, and V. Modi. Adjoint operator approach to shape design for internal incompressible ﬂow. In G.S.
Dulikravich, editor, Proceedings of the 3rd International Conference on Inverse Design and Optimization in Engineering Sciences,
pages 391–404, 1991.
[14] J.C. Huan and V. Modi. Optimum design for drag minimizing bodies in incompressible ﬂow. Inverse Probl. Engrg., 1:1–25, 1994.
[15] M. Desai and K. Ito. Optimal controls of Navier–Stokes equations. SIAM J. Control Optim., 32(5):1428–1446, 1994.
[16] N.A. Pierce and M.B. Giles. Preconditioning compressible ﬂow calculations on stretched meshes. AIAA Paper 960889, 34th
Aerospace Sciences Meeting and Exhibit, Reno, NV, 1996.
[17] N.A. Pierce and M.B. Giles. Preconditioned multigrid methods for compressible ﬂowcalculations on stretched meshes. Submitted
to J. Comput. Phys. April 1996.
[18] A. Jameson. Optimum aerodynamic design using control theory. Comput. Fluid Dynamics Review, pages 495–528, 1995.
[19] L. Martinelli and A. Jameson. Validation of a multigrid method for the Reynolds averaged equations. AIAA Paper 880414, 1988.
[20] A. Jameson, W. Schmidt, and E. Turkel. Numerical solution of the Euler equations by ﬁnite volume methods using Runge–Kutta
time stepping schemes. AIAA Paper 811259, 1981.
[21] A. Jameson. Analysis and design of numerical schemes for gas dynamics, 1, artiﬁcial diffusion, upwind biasing, limiters and their
effect on multigrid convergence. Internat. J. Comput. Fluid Dynamics, 4:171–218, 1995.
[22] P.L. Roe. Approximate Riemann solvers, parameter vectors, and difference schemes. J. Comp. Phys., 43:357–372, 1981.
[23] A. Jameson. Analysis and design of numerical schemes for gas dynamics 2, artiﬁcial diffusion and discrete shock structure.
Internat. J. Comp. Fluid Dynamics., 5:1–38, 1995.
[24] L. Martinelli. Calculations of viscous ﬂows with a multigrid method. Thesis Princeton University, May 1987.
[25] A. Jameson. Multigrid algorithms for compressible ﬂow calculations. In W. Hackbusch and U. Trottenberg, editors, Proceedings
of the 2nd European Conference on Multigrid Methods, Cologne, 1985, pages 166–201. Lecture Notes in Mathematics, Volume
1228. SpringerVerlag, Berlin, 1986.
[26] A. Jameson. Reengineering the design process through computation. AIAA Paper 970641, January 1997.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.